Arrakis

Serve any model from Hugging Face on a machine of the person's, like a hosted inference API but on their own GPU: vLLM or SGLang for text models, vLLM-Omni or SGLang for image and video models (FLUX, Qwen-Image, Wan), vLLM-Omni for speech and music, llama.cpp for a GGUF file (owner/repo:Q4_K_M). extra_args takes the person's own engine options as they'd paste them. COSTS MONEY like rent_gpu: the machine's hourly price, paid an hour at a time from when it runs until terminate_instance or termination_hours; worst_case_usd is the most it can cost. The model's size is read from Hugging Face and, without node_id, the best-ranked clean machine that holds it is rented on as few GPUs as it allows. A gated or private model needs hf_token (read access is enough; it goes to the machine only and Arrakis keeps none). Returns the instance, the endpoint (an OpenAI-compatible base URL once RUNNING, and its api_key) and ssh_command: the person also gets root SSH. The engine downloads the model first, so it answers a few minutes after RUNNING: get_instance shows endpoint_ready. Same idempotency_key returns the same machine.

The account balance (what it can pay now, what the running machines cost per hour, when the money runs out, whether rentals are paused and whether auto-recharge tops it up from a saved card) and this token's spend cap: the limit, what is committed, what is left, the longest rental allowed and when the token expires. Call it before long jobs and whenever an alert mentions the balance. Free; changes nothing.

Events on your machines, newest first: machine.ready (with its SSH command), machine.ended (reason says why: evicted when its host took it back or stopped working, otherwise stopped, removal_time, out_of_funds, paused or failed) and machine.notice (its host will take it back or restart it at notice_at, stopped answering, or it failed). Check it when a machine stops answering, and after any alert on a tool result. Free; changes nothing.

One machine now: state (PENDING or PROVISIONING while it boots, which is free; RUNNING when ssh_command works; then TERMINATED or FAILED), price per hour, ssh_command, removed_by, any notice from its host (host_notice and host_notice_at, host_down_since) and, once it ended, ended_reason (evicted when its host took it back). After rent_gpu, call it every 30-60 seconds until RUNNING. Free; changes nothing.

Live market summary by GPU family: cheapest price per GPU-hour, GPUs available and machines. Free; needs no token; changes nothing.

For partners (agents and people who bring customers to Arrakis): your partner codes with the link and a message to share, what each earns you (a share of what the customers you bring pay), the customers your codes brought (each with what they paid and earned you, and your link tag; never who they are), what you have earned, what is pending until those payments reach Arrakis or held before a cash-out, what you can cash out or move to your balance, any sales target with its progress, and whether the partner program is open to join. Free; changes nothing. Wherever you recommend Arrakis, say that you earn from referrals (the partner terms). Needs a token with the billing scope.

List, add or take off the SSH keys that can log in to a running machine. With only instance_id: the keys on it (fingerprint, comment, the saved key's name). add_ssh_key_ids / add_public_keys: more keys can log in (a teammate's, a new laptop's), at once. remove_fingerprints: those keys can't log in any more (rotate a key: add the new one and remove the old one in one call); taking keys off takes up to half a minute, and at least one key must stay. Free: it changes no billing. Keys are never returned, only fingerprints.

Join the Arrakis partner program and get your own partner code, link and a message to share. The account earns a share of what the customers it brings pay, as cash (paid through Stripe to the person or company you act for, once they set up payouts in the console) or Arrakis credit you can spend on GPUs. Only join if the person or company you act for agrees: joining accepts the partner terms (terms_url from get_partner_earnings) for them. Pass the terms_version you read. Free; joining again returns the same code. Add ?tag=<label> to your link to tell your own channels apart. Needs a token with the billing scope.

Images a machine can boot, for rent_gpu's template_id: the Arrakis images (PyTorch, CUDA toolkit, Ubuntu; SSH, login banner and tools built in) and the account's saved ones, with download size. Any other public image works through rent_gpu's image. Free; changes nothing.

Every machine on the account with its state, price per hour, SSH command, removal time and any notice from its host. Check it at the start of a session: a machine left running keeps billing until its removal time. Free; changes nothing.

What serving a Hugging Face model takes, before anything is rented: the GPU memory it needs and the machines it fits on right now, best value first, each with its hourly price and a rough speed. Free; rents and changes nothing. Call it first whenever the person wants a model running (to chat with, call from code, generate images, video or speech, or run evals against), tell them the price, then deploy_model with the option's node_id and gpu_count.

Rent one machine found by search_gpus, for work you run yourself over SSH (training, fine-tuning, batch jobs). To serve a model, use plan_deploy then deploy_model instead: one call sets up the machine, the engine and an OpenAI-compatible endpoint. COSTS MONEY: price per GPU-hour x GPUs, paid an hour at a time from when it starts running (booting is free) until terminate_instance or termination_hours, whichever comes first; worst_case_usd in the result is the most it can cost, and balance_covers_hours says how many hours the balance pays. It rents every GPU on the machine unless gpu_count says fewer (only some hosts allow that). Pass max_price_per_gpu_hour, normally the price search_gpus showed: if the live price is higher nothing is rented and the error gives the new price. Spot machines can be taken back by their host at any time. Boots the Arrakis PyTorch image unless template_id (list_images) or a public image is given; uses ssh_public_key, else ssh_key_id, else the account's first SSH key. Returns instance_id, state, the locked price, worst_case_usd and what to do next. Calling again with the same idempotency_key, or the same node_id and name within 10 minutes, returns the same machine instead of renting a second; use a new name for a second machine.

Change when a machine is removed automatically (ISO 8601, UTC). It bills until then, and removal deletes the machine and its disk. Use it to end a rental early when the job will finish sooner, or, with a token that has no spend cap, to extend a long run. A token with a spend cap can only move removal earlier.

Search GPU machines you can rent now at live prices: B300, B200, H200, RTX PRO 6000, RTX 5090, H100 and more. Free; changes nothing. Use it before rent_gpu and to answer what a GPU costs right now. Rows come best first: what the GPU can do (capability, H100 = 1.0), its price and how reliable its host is, with badges saying why (Reliable host, NVLink, or a warning such as a host taking it back soon); sort=price lists the cheapest first. pick is the first row with the reason, cheapest the lowest whole-machine price. Each row has node_id (for rent_gpu), gpu, gpu_count (rent_gpu rents all of them unless gpu_count says fewer; max_gpus 1 finds single-GPU machines), vram_gb per GPU, price_per_gpu_hour_usd, price_per_hour_usd for the whole machine, CPUs (and cpu, the model), RAM, disk, location, tier and network. Badges also say when renters rate a host low, it often fails to start or its GPUs are slow next to the same GPU elsewhere. Tiers: on_demand runs until its removal time; spot costs less but its host can take it back at any time, so use it only for work that saves checkpoints off the machine. Prices can move within a minute: pass the row's price to rent_gpu as max_price_per_gpu_hour. For multi-machine training add min_nic_gbps 400 and min_nics_per_gpu 1; null network fields were not reported.

Stop a machine now and delete it with everything on its disk; billing stops (the hour already started is paid). Copy results off first, for example with scp or rsync over its ssh_command. Cannot be undone; terminating a machine that already ended changes nothing.